Opal
ActiGraph wGT3X-BT
Sampling a high frequency component (red) at a low sampling rate will cause the high frequency component to look like a low frequency component (blue).
This could get mixed up with real low-frequency components.
Additionally, when we then low-pass the data, it might become distorted.
This might not be such a large problem in gait data.
There is a lot of power concentrated to below 15 Hz.
See articles:
https://alkvi.github.io/imu_comparison/#frequency-components-of-gait
The MobGap implementation uses the Madgwick filter to estimate global frame orientation, to better estimate stride length and walking speed.
This requires a gyroscope (and preferably a magnetometer).
The ActiGraph does not have this.
However, to at least counteract variations in ActiGraph placement, the vertical axis could be aligned with gravity before each gait sequence, if we detect a stable period before the gait sequence.
For example using something like Moe-Nilssen 1998.
The same MobGap pipeline was run on both lumbar Opal and hip ActiGraph data.
Here is the same snippet of gait data from the Opal and ActiGraph bandpassed between 0.15 Hz and 3.14 Hz, as used in initial contact detection.
There does not seem to be issues with high frequency components, at least during stable gait such as this example.
This a comparison of detected gait sequences and initial contacts for an approximately 8 minute walk outside including variations in pace, turning, some stopping. There is a difference of 4 found initial contacts (772 vs 768).
This comparison of stride length without orientation estimation (left) vs with orientation estimation (right) illustrates a slight systematic error for the ActiGraph, but subject parameters can clearly be delineated. Values are time binned (the _per_sec_ MobGap outputs).